TL;DR
This paper introduces "Andromaly," a lightweight, on-device Host-based Intrusion Detection System (HIDS) for Android. It uses machine learning to monitor system metrics and classify application behavior as normal or malicious, demonstrating effectiveness in differentiating between benign app categories.
What Problem Does This Paper Address?
Smartphones, particularly Android devices, have evolved into powerful, networked computers, making them prime targets for malware. Traditional security solutions like signature-based antivirus are often insufficient due to resource constraints (CPU, battery) and the rapid emergence of new malware. This paper addresses the challenge of detecting unknown malware on Android devices by proposing a behavioral-based detection framework that runs directly on the device, minimizing resource overhead.
Methods
The core of Andromaly is a Host-based Intrusion Detection System (HIDS) that continuously monitors various system metrics and events on the Android device. These metrics are then fed into machine learning classifiers to distinguish between normal (benign) and abnormal (malicious) behavior. The framework comprises several components:
- Feature Extractors: Collect system metrics (e.g., CPU usage, network traffic, battery level, running processes) from the Android framework and Linux kernel.
- Feature Manager: Manages the extraction process, triggering feature collection at defined intervals and potentially performing pre-processing.
- Processors: These are the analysis and detection units. They can be rule-based, knowledge-based, or, as explored in this paper, machine learning classifiers (e.g., Decision Trees, Naïve Bayes, Logistic Regression).
- Threat Weighting Unit (TWU): Aggregates the results from multiple processors using ensemble methods to produce a final threat assessment.
- Alert Manager: Smooths the threat assessment over time to avoid false alarms and triggers actions (e.g., uninstalling an app, displaying a notification).
The paper also investigates feature selection techniques (Information Gain, Chi-Square, Fisher Score) to identify the most discriminative features and reduce computational load, crucial for mobile environments.
Datasets
The authors created their own dataset due to the lack of standard benchmarks for this specific task. They collected:
- 23 game applications and 20 tool applications.
- These applications were installed on five different Android devices (HTC G1) to capture variations in usage patterns.
- For each application, the HIDS collected feature vectors every 2 seconds for 10 minutes, resulting in approximately 300 feature vectors per application per device.
- The dataset was labeled with the true class: 'game' or 'tool'.
- A total of 88 features were collected, categorized into Application Framework and Linux Kernel features.
Key Insights and Ideas
- Behavioral detection is feasible on Android: The successful differentiation between game and tool applications demonstrates that machine learning can effectively model and learn the behavior of Android applications.
- Lightweight detection is achievable: The framework's on-device nature and the effectiveness of using a small number of features (top 10-20) indicate that resource constraints on mobile devices can be met.
- Feature selection is crucial: Identifying the most discriminative features significantly improves detection efficiency and accuracy, making the system more practical for mobile deployment.
- Cross-device validity: The experiments showed similar performance across different devices, suggesting that the learned models are not overly dependent on specific device configurations or user behaviors, pointing to persistent application behaviors.
Key Results
The paper evaluated various combinations of classifiers and feature selection methods.
- Best Classifiers: Decision Tree (DT), Logistic Regression (LR), and Bayesian Networks (BN) consistently performed well in both experiments.
- Best Feature Selection & Number of Features:
- In Experiment I (training and testing on the same application types), InfoGain with the top 20 features yielded the best results, achieving an average accuracy of 0.877 and a False Positive Rate (FPR) of 0.129.
- In Experiment II (testing on application types not seen during training), Fisher Score with the top 10 features performed best, with an average accuracy of 0.750 and an FPR of 0.250.
- Top Features: Features like
Load_Avg_15_mins,Total_Entities,Running_Processes,Context_Switches, andSchedule_Callswere consistently ranked high across different methods and experiments, indicating their importance in distinguishing application behavior.
Limitations and Future Work
- Limited Malware: The study used game and tool applications to simulate detection due to the scarcity of actual Android malware at the time. This means the direct effectiveness against real malware is not fully demonstrated.
- Cross-device generalization: While similar performance was observed, further research is needed to understand how well classifiers trained on one set of devices can detect anomalies on entirely different devices.
- Temporal aspects: The current feature set is largely static. Future work could incorporate temporal information (e.g., changes in battery level over time, sequences of events) to capture more dynamic behavioral patterns.
- Application-specific detection: The current system monitors the whole system. Focusing on monitoring individual application processes could lead to more precise detection.
- Persistent anomalies: The framework could be enhanced to alert only when an anomaly persists over time, reducing transient false positives.
Summary updated